Y Combinator’s Paper Club turns to inference this week, making the case that for every FLOP spent training a model roughly six go to serving it each year, and that getting that half of the stack right has produced more than $220B in market cap over the past decade.
The session traces an inference request from the API down to the kernel, examines why the gap between the best and worst provider running the same open model exceeds the gap between open and closed models, and closes with SGLang and memory management for agents.
